Jul 13, 2023|

JD.com Introduces ChatRhino: Empowering Industry Innovations with an Advanced Large Language Model


by Yuchuan Wang and Vivian Yang

JD.com today unveiled its ChatRhino (Yanxi in Chinese) large language model (LLM) on its 2023 JDDiscovery tech summit, tailored to serve various industries. By combining 70% generalized data with 30% native intelligent supply chain data, JD’s latest AI model offers targeted solutions for real industry challenges across sectors such as retail, logistics, finance, health, and city. Building upon the success of the billion-parameter model K-PLUG launched in 2021 and the 10-billion-parameter model Vega introduced in 2022, JD’s ChatRhino sets a new benchmark as a 100-billion-parameter model.

JD.com has also established a comprehensive LLM tool chain. This includes the MaaS platform “ChatRhino AI Developing and Computing Platform”, the vector database “Vearch,” the upgraded hybrid multicloud OS “Cloud Ship,” the high-performance storage platform “Cloud Ocean,” and the hardware and software integrated virtualization engine “Jing Gang.” In addition to the powerful ChatRhino LLM, JD.com also announced the upgrade of the Youjia DaaS platform and ChatRhino Intelligent Service platform.

“JD’s large model evolution aligns with our relentless pursuit of technology, encompassing the pillars of cost-effectiveness, efficiency, user experience, trustworthiness, inclusiveness, and groundbreaking progress. Our pursuit of cost, efficiency, and experience are deeply ingrained in JD’s business philosophy, serving as the first principles of retail. Moreover, our commitment to trustworthiness, inclusiveness, and breakthrough signifies our dedication to leveraging technology to benefit industries and society as a whole,” said Sandy Xu, CEO of JD.com.


An Industry-Savvy Large Model Trained for Real-World Applications

JD’s extensive industry coverage provides ChatRhino with unparalleled advantages. The model can accumulate tens of billions of high-quality interactive data and know-how derived from JD’s various industry scenarios every year. Additionally, ChatRhino possesses exceptional industry generalization capabilities and ensures secure usage for clients both on-cloud and off-cloud.

Peng Cao, Chair of JD.com’s Technology Committee and President of JD Cloud, said, “The digital intelligent supply chain trains industrial large models, while large models leverage supply chains to go deep into industries.”

To deploy ChatRhino’s applications in real-world business scenarios, Dr. Xiaodong He, Director of JD Explore Academy and President of JD Technology’s Intelligent Services and Products Division, outlined a three-step approach:

  • Step 1: By now, JD.com has built a generative large model based on practices within JD.com’s diverse business operations.
  • Step 2: Enhance and iterate the model through highly complex industrial scenarios, creating robust services for industry use by end of the year.
  • Step 3: Fully open up the capabilities of the model for serious commercial applications in early 2024.

Training Tool Kits for Rapid Model Deployment in Industries

JD’s ChatRhino AI Developing and Computing Platform offers tailored solutions for clients’ large models and industry-specific application development. With over 100 training and inference optimization tools, the platform facilitates the seamless transformation of generative models into specialized ones, eliminating the need for lengthy and resource-intensive procedures.

During the summit, JD showcased examples of rapid model deployment, including converting a generative model into a healthcare industry-specific model. What traditionally required a week and a team of over ten scientists can now be accomplished by just 1 or 2 algorithm engineers within minutes. The platform’s model accelerating and optimizing tools significantly reduce model inference costs by 90%.

JD Health’s Jingyi Qianxun (meaning asking doctors thousands of times) Large Model, based on the ChatRhino, utilizes rapid migration and learning of various medical and health scenarios to automatically deploy products and solutions. This technology serves as a robust foundation for telemedicine services.

Another notable demonstration showcased an AI marketing operation platform within the financial sector. Through simple conversations, users can effortlessly create comprehensive marketing campaigns, including operational strategies, personnel arrangement, quick campaign pages setup, bulk generation of marketing copy and materials, as well as digital delivery. Previously, achieving these results required coordination among several different teams, spanning product, R&D, algorithms, design, and analysts. However, now a single person can accomplish these tasks. Moreover, this solution significantly reduces human-computer interactions from over 2,000 to fewer than 50. This innovative approach has the potential to increase marketing production efficiency by a hundredfold.

In the e-commerce domain, the JD Cloud AIGC content marketing platform can now swiftly produce a diverse range of essential e-commerce visuals, captivating marketing posters, and detailed product images using just one product image. These assets cater to the demands of rapid store setup, product promotion, and other marketing requirements. With this advancement, the production cost of each visual set has been reduced by 90%, while the time required has been slashed from seven days to half a day.

If the ChatRhino AI developing and computing platform is the training center for JD’s LLM, the Vearch vector database is the bridge for the AI model to utilize data. Vearch can now support high-performance retrieval from tens of billions of data sources, with latency reduced to milliseconds and availability reaching 99.99%. It has already served over 100 large and medium-sized enterprise users. JD.com has used the vector database for large-scale pre-training of models, reducing inference costs by 80%.

JD.com has also made significant investments in cloud computing infrastructure to support its LLMs. In 2021, JD established Tianqin α, China’s first large-scale computing cluster based on the SuperPOD architecture, located in Chongqing. Tianqin α achieves a remarkable 6.2 times acceleration in inference speed and a 90% reduction in inference costs. It has become an indispensable pillar of computational support for technological innovation.


Upgraded Multi-Scenario Solutions Drive Industry Innovation

“Intelligentization takes root first in industries that are at the forefront of digitalization,” said Liqiang Gao, Vice President of JD.com and head of JD Technology’s Solution Center. In sectors such as retail, finance, cities, and logistics, JD has introduced innovative intelligent solutions, harnessing the power of large models to enhance service capabilities and drive transformative change.

JD Logistics launched the 3.0 version of Jinghui Intelligent Supply Chain Data Management Platform. With the support of ChatRhino, Jinghui’s performance in sales forecasting, inventory management and replenishment has been greatly improved, while helping users locate supply chain problems and offer solutions quickly through its interactive supply chain control tower.

JD.com CEO Sandy Xu said, “JD’s journey in developing large language models is closely intertwined with our industry partners. JD is committed to forging strong collaborations with industry partners to continually contribute to the high-quality development of industries through technology innovations.”

JD’s commitment of technology centers around instilling reassurance and trust in its industry partners, particularly in the realm of AI represented by large models. This empowers partners to confidently leverage technology, enabling its transcendence beyond the confines of laboratories and corporate boundaries to create value within industrial scenarios.


(yuchuan.wang@jd.com; vivian.yang@jd.com)